The contents of the problem set for assignment 1 are here with the PDF of the assignment also included in this repository as assignment1.pdf
This uses python 2.7. Run the excercises by using python q{1,2,3}_{problem}.py, basically execute any code by passing the file to the python interpreter.
Some of the functionality has tests but some is still missing.
I think there is a large chunk of q3_word2vec.py that can benefit from sanity tests.
This assignment benefits from using GPUs so there are instructions on spinning up a spot instance on AWS. I use a p2.8xlarge instance which as of June 04th $7.20 per hour. Spot instance makes it possible to run this same instance type for $1.3 per hour.
To make a request for a spot instance run ./aws_setup/setup.sh currently this
assumes my laptop and uses an AWS profile that I had setup for my personal AWS
account. You would need to change things around and this probably won't travel
as well because the AWS AMI that I am using is in my own account. This doesn't
travel well at the moment.
config.json is a saved file from my first spot instance request using the AWS management console. The saved config.json and setup.sh scripts in the aws_setup folder make it much easier for me to re-request an instance next time I want to run a training.
Once instance is up there will be a cs224 folder in the home folder.
git pull origin master and you have the latest code. Execute with python q3_RNNLM.py
./aws_setup contains support scripts for creatin AWS spot instance infrastructure
source ./aws_setup/commands.sh to have some basic commands that perform:
describe the environment
setup a new spot instance
login to the instance
terminate it when you are done;
When running N multiple instances choose which one to login to by providing an
instance number from 0-N.
./aws_setup/login.sh 0 to login to the 1st instance
The best executions were trained in 90 minutes on p2.xlarge and had accuracy of .745 on validation.
Updating the embed size up to 2800 from the starting 35 did little to change the resulting accuracy.
The l2 regularization was as it should be super important in getting good validation results.
The training set and the validation set accuracies were still wide apart, the training set accuracy would frequently be above %92
All of the results of running Assignment 3 and individual write ups are in assignment3/output and assignment3/README.md